A sparse-feature-based face detector

TL;DR: This work construct an over complete set of weak classifiers using linear support vector machine algorithm, and then select part of them using the AdaBoost algorithm and combine the selected weak classifier to form a strong classifier that can minimize the classification error directly.

Abstract: Local features and global features are two kinds of important statistical features used to distinguish faces from non-faces. They are both special cases of sparse features. A final classifier can be considered as a combination of a set of selected weak classifiers, and each weak classifier uses a sparse feature to classify samples. Motivated by this thought, we construct an over complete set of weak classifiers using linear support vector machine algorithm, and then select part of them using the AdaBoost algorithm and combine the selected weak classifiers to form a strong classifier. During the course of feature extraction and selection, our method can minimize the classification error directly, whereas most previous works cannot do this. The main difference from other methods is that the local features are learned from the training set instead of being arbitrarily defined. We applied our method to face detection. The test result shows that this method performs well.